SAGE: Agentic Framework for Interpretable and Clinically Translatable Computational Pathology Biomarker Discovery

Despite significant progress in computational pathology, many AI models remain black-box and difficult to interpret, posing a major barrier to clinical adoption due to limited transparency and explainability. This has motivated continued interest in engineered image-based biomarkers, which offer greater interpretability but are often proposed based on anecdotal evidence or fragmented prior literature rather than systematic biological validation. We introduce SAGE (Structured Agentic system for hypothesis Generation and Evaluation), an agentic AI system designed to identify interpretable, engineered pathology biomarkers by grounding them in biological evidence. SAGE integrates literature-anchored reasoning with multimodal data analysis to correlate image-derived features with molecular biomarkers, such as gene expression, and clinically relevant outcomes. By coordinating specialized agents for biological contextualization and empirical hypothesis validation, SAGE prioritizes transparent, biologically supported biomarkers and advances the clinical translation of computational pathology.

💡 Research Summary

The paper addresses a critical barrier in computational pathology: the lack of interpretability in most AI models, which hampers clinical adoption. To overcome this, the authors introduce SAGE (Structured Agentic system for hypothesis Generation and Evaluation), an agentic AI framework that discovers engineered, interpretable pathology biomarkers grounded in biological evidence.

SAGE’s pipeline consists of three major components. First, a domain‑specific biomedical knowledge graph (KG) is automatically constructed from 1,650 scientific sources (including 1,000 bladder‑cancer papers, 650 computational‑pathology papers, and three textbooks). Using GPT‑4o‑mini, the system extracts biomedical triples from filtered literature, retains those with confidence ≥ 0.5, normalizes relations via ontology rules, and merges redundant entities with BGE‑Large embeddings (cosine similarity threshold 0.9). The final KG contains 41,053 nodes and 56,338 edges, covering genes, pathways, diseases, phenotypes, clinical outcomes, and image‑derived concepts. Validation on a stratified random sample shows 99 % factual grounding and 100 % relation accuracy, with entity‑type precision improving from 64.6 % to 82.5 % after refinement.

Second, SAGE orchestrates a sequence of specialized agents in a fixed order, eliminating the need for a central supervisor and reducing stochastic divergence. The agents are:

1. Path Generation – extracts raw image features (color, texture, morphology) from whole‑slide images (WSIs).
2. Ontologist – traverses the KG to uncover non‑obvious biological links between image features and molecular entities.
3. Scientist – translates these links into natural‑language hypotheses that explicitly state a putative relationship (e.g., “high TLS density correlates with CXCL13 expression”).
4. Hypothesis Expansion – generates variations and refinements of the base hypothesis to explore a broader hypothesis space.
5. Novelty Critic – compares each hypothesis against existing literature to assess scientific novelty and avoid rediscovery.
6. Feasibility Agent – evaluates data availability, statistical power, and clinical relevance, flagging hypotheses that are testable with the current multimodal dataset.

Third, the Coding Agent automatically synthesizes executable R/Python scripts that implement the statistical analyses required to test each hypothesis. It integrates WSI‑derived quantitative metrics with molecular data (e.g., RNA‑seq, gene‑expression panels) and runs appropriate models such as Cox proportional hazards, logistic regression, or correlation analyses. The Summary Agent then aggregates results, producing clinician‑friendly tables, plots, and concise narrative statements.

The authors demonstrate SAGE on a bladder‑cancer use case. The Ontologist discovers a link between tertiary lymphoid structures (TLS) visible in histology and the chemokine genes CXCL13 and LTB. The Scientist formulates the hypothesis that “regions with high TLS density exhibit elevated CXCL13 expression, which predicts improved overall survival.” The Coding Agent validates this by quantifying TLS density on WSIs, aligning it with patient‑level RNA‑seq data, and performing survival analysis. Results show a statistically significant positive association (p < 0.01) between TLS density, CXCL13 expression, and longer survival, confirming the hypothesis and providing an interpretable, biologically grounded biomarker.

Key contributions highlighted by the authors include:

• Biologically grounded agentic design that integrates literature‑derived knowledge with multimodal data, ensuring each hypothesis has a mechanistic rationale.
• Multi‑path ontological reasoning that captures diverse, plausible connections across imaging, genomics, and clinical domains.
• Debate‑based novelty assessment via the Novelty Critic, offering transparent evaluation of scientific originality.
• End‑to‑end validation pipeline where hypotheses are automatically translated into executable analyses, dramatically reducing manual coding effort.

Limitations are acknowledged. The KG construction relies on automated triple extraction, which can introduce noise; the current evaluation is limited to bladder cancer, so generalizability to other tumor types remains to be proven; and the multi‑agent workflow, while structured, can be complex to debug. Future work is proposed to incorporate expert‑in‑the‑loop KG refinement, expand to multi‑cancer cohorts, and develop richer provenance tracking for each agent’s decisions.

In summary, SAGE represents a significant step toward interpretable, clinically translatable computational pathology. By marrying large‑language‑model reasoning, structured biomedical knowledge graphs, and automated statistical validation, it offers a reproducible pathway from image‑derived features to mechanistically justified biomarkers, potentially accelerating the adoption of AI tools in routine pathology practice.